SYNCRETISM AS PF-REPAIR: THE CASE OF SE-INSERTION IN SPANISH

0 SYNCRETISM AS PF-REPAIR: THE CASE OF SE-INSERTION IN SPANISH Mercedes Pujalte and Andrés Saab 4 ABSTRACT Purpose This paper focuses on the mechanism that determines the distribution of clitic insertion (CL-insertion) in a set of well-known argument alternations in Spanish (reflexives, anticausatives, and passives/ impersonals). Approach We claim that a clitic is inserted post-syntactically when v [EXT ARG] is not merged with a DP in the syntax. The absence of a Spec,vP is due to the lack of j-features on C or v. By adapting Müller s (200) proposal that Merge is induced by subcategorization features, we argue that these features must be discharged at the morphological level for well-formedness conditions. If some of these features are not satisfied in the syntax, CL-insertion applies as a repair strategy to remove the surviving subcategorization feature. This is precisely the case in the empirical domain with which we are concerned: a v [EXT ARG] that is not merged with any DP in the syntax induces CL-insertion at PF as a last resort operation. Such an operation is subject to strict locality conditions that determine morphological visibility for cyclic computation (Marvin, 2002; Embick, 200). The End of Argument Structure? Syntax and Semantics, Volume 8 Copyright r 202 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 002-46/doi:0.08/S002-46(202)00000802

0 Mercedes Pujalte and Andrés Saab Findings We will show that a restrictive theory of CL-insertion not only derives the basic distribution of some argument alternations, but also accounts for why CL-insertion can rescue external arguments but not internal ones. Theoretical implications If our approach is on the right track, it follows that the notion of argument structure should be considered as an epiphenomenon and not as primitive of grammatical theory. Keywords: CL-insertion, argument alternation, subcategorization, feature inheritance, Distributed Morphology, Spanish. INTRODUCTION In this paper we propose a novel analysis for the problem of SE syncretism AU : in Spanish. As is well known, the third person clitic pronoun SE (or the corresponding agreeing clitic me, te, nos, etc.) occurs in a number of contexts such as anticausative (), passive (2), impersonal (), and reflexive constructions (4), among others. () Se hundió el barco con la tormenta. SE sank.sg the ship with the storm The ship sank with the storm. (2) Se cerraron las puertas para bloquear la salida. SE closed.pl the doors for to.block the exit The doors were closed in order to block the exit. () Nunca se castiga a los culpables. never SE punishes to.acc the culprits No one ever punishes the culprits. (4) Juan se lava. Juan SE washes Juan washes himself. We explore the hypothesis that the syncretism in () (4) follows from the PF condition in (): 4 As a PF condition, we expect to find variation across languages not only in the way in which () is satisfied but also in the way it is formulated. So it is a logical option that the reference to v [EXT ARG] is part of a parameter.

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 2 4 () At PF a clitic has to be inserted when v [EXT ARG] does not have a specifier. 2 Before illustrating the details of our proposal, let us clarify (). First, v [EXT ARG] AU :2 is shorthand for a little v that licenses external arguments (EA). Second, the claim that v [EXT ARG] does not have a specifier means that a transitive v is not syntactically merged with a thematic DP in the syntax, as has been proposed by Embick (2004) and several subsequent works (see especially Schäfer, 2008 for a detailed analysis). In this sense, clitic insertion (CL-insertion) is not a consequence of a valence-reduction operation like decausativization or reflexivization (contra Chierchia, 2004; Levin & Rappaport-Hovav, ; Reinhart, 2002; Reinhart & Siloni, 200 for a mixed approach), but a rescue operation taking place in the PF branch. In this respect, our proposal is in consonance with other syntactic approaches to the problem of valence reduction, as is formulated in Distributed Morphology (see Alexiadou & Anagnostopoulou, 2004; Alexiadou, Anagnostopoulou & Schäfer, 2006; Embick,, 2004; Folli & Harley, 200; Schäfer, 2008, among others). However, one of the novelties in () is the role that PF plays in deriving syncretism in these contexts. We argue that the clitic SE (or the relevant agreeing clitics) occurring in () (4) is only inserted postsyntactically because of the need to satisfy some selectional property of v [Ext Arg]. If () is on the right track, we have to answer at least the following questions: (A) (B) Under which particular syntactic scenario is the absence of an EA in the syntax allowed? How is CL-insertion implemented at PF? As an answer to question (A), we propose that the inflectional system associated with the C-Tand v-o complexes (see Chomsky, 200, 2008) plays a crucial role in deriving both the syntactic reduction effects and some of their thematic properties. According to this idea, what determines the argument reduction effects in () (4) is the set of possible j-feature combinations on 2 It should be noted that this late insertion approach to SE constructions resembles the late insertion approach of expletives proposed by Vukić (200) for languages like English. The reader is referred to this work and the references therein for more details. An anonymous reviewer questions the PF character of categorial features and claims that they are syntactic diacritics for semantic relations. However, as the reviewer acknowledges, this semantic connection between selector and selectee is only one of the aspects of the syntax of categorial features. As is well known, the status of such features is indeed controversial at least since Grimshaw () and Pesetsky (82). The debate centers on whether or not c-selection could be eliminated in favor of s-selection. As far as we know, this reductionist move has presented more problems than solutions. Therefore, there is a purely formal residue of categorial features that affects the syntax-pf connection. This paper, as much of current research in the minimalist framework, centers on this aspect of the problem. As far as the semantic import of categorial features is concerned, we do not have an explicit interpretative mechanism to offer. We only note below that the semantic consequence of a non-discharged categorial feature seems to be directly connected to some particular thematic interpretations of reflexives/reciprocals and to the semantics of arbitrary readings in general.

2 Mercedes Pujalte and Andrés Saab 4 C and v. Thus, we will show that the feature inheritance mechanism (Chomsky, 200, 2008) and the possible j-feature combinations on the core functional categories are closely linked to certain argument structure effects. In our system, v might lack an EA if either the v-o complex or the C-T complex is associated with no j-features from the numeration. The answer to question (B) is the heart of our proposal. We will suggest that there is an explicit mechanism for CL-insertion at PF, whose consequences go beyond the problem of SE syncretism in Spanish. In a nutshell, assuming that Merge is motivated by the need to satisfy categorial features (or structurebuilding features in Müller s, 200 terms), the absence of a DP in the EA position of v causes a crash at PF, unless PF has some way to repair this subcategorization problem. We propose that CL-insertion is precisely a way to circumvent this failure. If this idea is on the right track, we must conclude that clitics have to be divided into at least two classes: (i) syntactic clitics (i.e., arguments introduced in the syntax) and (ii) morphological clitics. As we will see, morphological clitics have expletive-like properties, in the sense that they lack interpretable/valued j-features, do not have a thematic role, and need an associate to value their features. At the same time, morphological clitics occupy specific positions in the clitic cluster: they always occur above syntactic clitics, at least in Spanish. In our system, this fact is accounted for under the following two hypotheses: (i) Agree is implemented in the morphology using the same Probe-Goal relation as in the syntax, with the difference that it applies only in the domain of heads (or complex heads) and (ii) morphological derivation is subject to phase conditions, as argued in Marvin (2002) and Embick (200), among others. We show how the combination of hypotheses (i) and (ii) explains the mechanism of CL-insertion, and also why CL-insertion, as a repair strategy, is not attested in object position. The paper is organized as follows. In Section 2, we describe the basic distribution in the vp domain of the SE constructions given in () (4). In Section, we present our basic assumptions about the nature of the inflectional and subcategorization features that trigger the syntactic operations of Agree and Merge. In Section 4, we propose an analysis of the syntax of SE constructions and show that they share the absence of a Spec,vP as their main property. In Section, we propose an explicit mechanism for CL-insertion at PF under the cyclic approach to morphology proposed by Marvin (2002) and Embick (200). We observe that CL-insertion can apply to EA positions but not to internal ones: that is, there are no expletive-like elements in object position. Section 6 summarizes the main findings of the paper. 2. THE DISTRIBUTION OF SE CONSTRUCTIONS IN SPANISH In this section, we present our basic assumptions about the vp domain and describe the distribution of SE constructions in Spanish. The purpose of this

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 2 section is to show that the main property that all SE constructions share is the absence of an EA. Following Folli and Harley (200), we assume that little vs introducing EAs are associated with [DO] or [CAUSE] features, whereas pure unaccusative structures are associated with a [BECOME] feature. The [DO] and [CAUSE] features partially determine the thematic interpretation of EAs. Compare a pure unaccusative verb like (6) with transitive verbs that have agents () or causes (8) as their EA: (6) Juan murió. Juan died.sg () Juan asesinó a Pedro Juan murdered.sg to.acc Pedro (8) El calor quemó el pasto. the heat burned.sg the grass In () and (0) the simplified structures corresponding to the sentences in (6) (8) are shown. On the one hand, () represents the vp structure of a pure unaccusative like morir to die (cf. (6) above). In this case, the feature associated with v is [BECOME], hence no EA is licensed and the internal argument (IA) of the Root becomes the syntactic subject. On the other hand, (0) represents the sentences in () and (8). Here, v licenses an EA that will be interpreted as an agent or as a cause, depending on the feature associated to v. () (0) 4 Let us assume that the intransitive version of (8) (e.g., El pasto se quemó AU : Lit: the grass SE burned.sg ) is not represented as in (). Instead, this intransitive sentence has essentially the same underlying structure as its transitive counterpart (see Alexiadou & Anagnostopoulou, 2004; Alexiadou et al., 2006; Embick, 2004; Schäfer, 2008, among others). If this is the case, the sentence in (), which might receive either a passive or an anticausative reading, follows from the selection of a v [DO] or a v [CAUSE]. If v [DO] is selected, the passive reading is forced, but when v [CAUSE] is selected, the predicate receives the anticausative interpretation.

4 Mercedes Pujalte and Andrés Saab () Se hundieron los barcos. SE sank.pl the ships Reading : The ships were sunk. Reading 2: The ships sank. There are several tests that distinguish the agentive and the causative readings. First, the anticausative interpretation is impossible with pure agentive verbs like construir to build. Thus, in (2b) the only available reading is the passive one. Second, passive SE, but crucially not anticausative SE, allows agentoriented modifiers (). Third, cause-oriented modifiers are only possible with anticausatives, as illustrated in (4): (2) a. Juan construyó la casa. Juan built.sg the house b. La casa se construyó. (passive OK, anticausative) the house SE built.sg The house was built. () El barco se hundió a propósito (passive OK, anticausative) the ship SE sank.sg to purpose The ship was sunk on purpose. (4) El barco se hundió con la tormenta ( passive, anticausative OK) the ship SE sank.sg with the storm The ship sank with the storm. Another basic difference between anticausative and passive SE constructions is related to the inflectional properties of the subject. Passive SE is subject to a person restriction that prevents the occurrence of strong personal pronouns or proper nouns as subjects (see Burzio, 86; Cinque, 88; Di Tullio, 200; D Alessandro, 200; López, 200; Taraldsen, ; among many others): 4 () a. Se encontraron cadáveres. SE found.pl bodies Bodies were found. b. Se encontró Juan/él. SE found.sg Juan/he c. Me encontré yo. ME found.sg I 4 4 This is the case in Spanish, but not in other Romance languages. In Italian, for example, proper names are allowed in passive SI constructions (Roberta D Alessandro p.c.).

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 2 In cases such as (b) (c), the only available interpretation is that of a reflexive. As is shown in Section 4, the difference between passive and reflexive SE depends on specific inflectional features: the inclusion of person features in the vp automatically yields the reflexive meaning. That is, passives and reflexives share the same type of vp with a v [DO], as illustrated in () (see Marantz, 84 among many others for an unaccusative analysis of reflexives ). Anticausatives, in turn, have the same structure but with a v [CAUSE] (6). Anticausative (6) () Passive SE/reflexives 4 The crucial property that both (6) and () share is the lack of a specifier for the v node. We would like to argue that this property is also present in the impersonal SE: a construction that has resisted unification approaches in the Romance literature (see D Alessandro, 200 for references and thorough discussion): (8) a. Se castiga a los culpables. SE punishes to.acc the culprits One punishes the culprits. b. Se trabaja duro. SE works hard One works hard. Impersonal SE has an arbitrary interpretation and, on a par with the passive, allows agent-oriented modifiers (). () Se persigue a los evasores para cobrar los impuestos. SE chases to.acc the evaders for to.collect the taxes. One chases the evaders to collect the taxes. One property that clearly distinguishes impersonal and passive SE is Case: the IA of passive SE values nominative, whereas the IA of impersonals Marantz s approach has been criticized in a number of different works, such as Reinhart and Siloni (200) and Labelle (2008). Due to space limitations, we cannot address this criticism here, but we believe that the arguments presented in favor of or against the unaccusative analysis are quite controversial.

6 Mercedes Pujalte and Andrés Saab 4 values accusative, as the pronominalization by accusative pronouns in (20) shows: 6 (20) En este país se los castiga. in this country SE Cl.acc.pl punishes In this country, one punishes them. Finally, in impersonal SE contexts the verb occurs invariably in third singular person: () a. Se castiga a los culpables. SE punishes to.acc the culprits b. Se castigan a los culpables. SE punish.pl to.acc the culprits We assume that the basic vp domain of impersonal SE is the same as that of reflexive/passive SE when it is transitive (see 8a). For cases like (8b), the underlying structure is the same as unergatives (parentheses in (22) only for convenience): (22) Impersonal The structure in (22) still does not account for some of the properties of impersonal SE mentioned so far (default agreement, for instance). Other characteristics of impersonal SE, such as its distribution in non-finite contexts (see Cinque, 88 and Section ), do not follow from (22), either. These problems will be discussed at length in the following sections. In summary, all the SE constructions introduced so far share the property of not having an EA merged in Spec,vP. In Section 4, we present an analysis of the syntax of SE constructions and, in Section, we propose an explicit morphological mechanism for CL-insertion at PF. However, before addressing these topics, we have to present our basic assumptions about inflectional and subcategorization features. 6 It should be noted that pronominalization by an accusative pronoun is allowed in River Plate Spanish but not in other Spanish dialects, like Mexican Spanish, where a dative clitic is used even though this dialect is not leísta (e.g., Se les castiga Lit: SE them.dat punishes ). Nevertheless, Ordóñez and Treviño (200) have shown that this dative clitic behaves as a direct object in all the relevant respects.

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 2. SOME ASSUMPTIONS: FEATURE INHERITANCE AND SUBCATEGORIZATION We assume the Agree system (Chomsky, 2000, 200) and a version of the feature inheritance mechanism recently suggested by Chomsky (200, 2008). According to Chomsky, T has no j-features; instead, T inherits them from C and, as a consequence, a C-T complex is formed. The same is postulated for the Root: The OP (or VP in Chomsky s system) does not have j-features and must inherit them from v, forming a v-o complex. In Chomsky s system, nothing prevents C or v from lacking j-features. In a situation where v has no j-features, v (and then the Root) inherits its j-features from C, forming a new complex C-T-v-O. In this context, the IA values nominative case. The opposite situation is also possible: C lacks j-features, but v is j-complete. In this case, no inheritance is available under the assumption that inheritance cannot take place in a bottom-up fashion. In (), these three possibilities are respectively illustrated: () a. b. c. 4 It is worth noting that there is a correlation between the feature composition of C/v and the number of potential DP arguments that can be merged in the derivation. Consider for instance the structure in (b) where v lacks j-features and a C-T-v-O complex is formed. Here, the occurrence of two DPs with unvalued Case features should produce a crash, because at least one of the two DPs would not have valued its Case feature or, depending on some assumptions concerning morphological Case, two DPs would be valued with nominative Case, an illegitimate output at PF. In other words, inheritance from C-T to v-o is restricted to the presence of one argument DP with structural Case. Furthermore, notice that if Probe-Goal dependencies can only be established within the complement domain of the Probe (Chomsky, 2000 and subsequent work), it follows that whenever C-T-v-O is formed, the only available DP in the derivation should be merged as an IA. Therefore, the impossibility of inheritance from C-T to v whenever Spec,vP is filled with an Assuming that in a simple transitive sentence only C or v can be inherently specified for [uj], the remaining option to explore is one where neither C nor v is specified with [uj]. This could be the case with pure impersonal sentences or weather predicates (e.g., llover, to rain ). We leave a careful discussion on this possibility as a matter of future research.

8 Mercedes Pujalte and Andrés Saab EA follows naturally: 8 an EA in Spec,vP is outside the search domain of v and cannot enter into an agreement relation with v (see Chomsky, 200, p. and Richards, 200 for more discussion). Of course, other syntactic contexts should prevent inheritance from C-T to v-o. It seems, for instance, that inheritance stops whenever a potential locus of inheritance is already specified with (in)complete j-features. Another situation that should be prevented is inheritance from C-T to v-o in unergative structures, otherwise the EA in Spec,vP will remain outside the search space of v-o for the reasons adduced by Chomsky (200) (see above). One option is that unergatives are underlying transitives with an incorporated object (Hale & Keyser, ). We can assume then that object incorporation and j-specification on v are in complementary distribution (maybe a controversial hypothesis, see Roberts, 200). We leave a careful exploration of all these issues for future research. With respect to the Merge operation, we assume with Müller (200) that Merge is triggered by an ordered set of subcategorization or structure-building features associated to functional heads. According to Müller, this kind of feature must be discharged by Merge in order to remove the subcategorization features from the syntactic computation. This approach presupposes that the deletion of subcategorization features is entirely effected in the syntax by virtue of syntactic Merge. This means that if some subcategorization feature is not discharged in the syntactic component, the derivation is immediately aborted. However, we depart from Müller s proposal in this respect. Instead we suggest that the satisfaction of subcategorization properties is evaluated at PF. We formulate this idea in (24): (24) At PF, every structure building feature must be discharged. 4 8 That is, unless this EA receives inherent Case or there is another functional head for valuing its Case feature. Quirky subject constructions in Icelandic and other languages could be an instance of this type. Thanks to an anonymous reviewer for pointing out this issue. The reviewer also wonders whether inheritance from C-T to v-o could be counter-cyclic, assuming category-defining v to be a phase head. There are several ways that can be pursued to deal with this potential problem. On the one hand, it could be the case that Agree is not constrained by phases. This is proposed in Bošković (200), where several empirical arguments are presented (see also Chomsky, 2008, p. 4). On the other hand, one can take the opposite direction and claim that feature inheritance is indeed the mechanism that makes Agree a very local operation, that is, phase-based operation (long distance agreement facts being an epiphenomenon). Assume with Chomsky (200) that valuation and transfer happen together (see also Richards, 200, 20 for discussion and references). If this is correct, inheritance from C-T to v-o is not a countercyclical operation, but is in consonance with a strict version of the Phase Impenetrability Condition. In a situation in which v has no inherent j-features, transfer of its complement has to wait until inheritance from C-T to v-o has applied (i.e., inheritance precedes Agree as suggested by Chomsky, 200, p. 22). We will leave this issue for future research.

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 2 Let us illustrate the basic mechanics of our adaptation of Müller s system with a simple transitive sentence like (). We focus only on the part of the derivation that corresponds to the vp domain. () John read the book. The two relevant heads are v and O which have the following structure-building features associated to them: O [D] and v [O]o[D], whereoindicates precedence in the order of Merge. Once O [D] enters the derivation, its D feature triggers Merge with a DP, forming [ OP O [D] DP]. The feature D on O is discharged by this instance of Merge. Then, v [O]o[D] is selected and its O feature (the first in the list) is discharged by merging v with the OP that is present in the working space. Finally, the D feature on v activates another instance of Merge with a DP. As a result, all the subcategorization features on O and v are correctly discharged and the derivation proceeds. In (26), we represent the vp with all the subcategorization features discharged. (26) In order to understand the difference between our assumptions and Müller s, consider a hypothetical situation where there is no DP discharging the D feature on v. 0 Under this circumstance, the derivation is aborted following Müller s approach, but not under our interface condition in (24). Assume now that in (26) T is introduced in the derivation with a v feature. As far as the syntactic derivation is concerned, T and vp can in principle be merged even if the head of the vp has its D feature not discharged: () Once this structure is sent to PF, there are two possibilities: (i) the nondischarged D feature on v produces a derivation crash or (ii) PF can provide some mechanism to rescue the derivation. The abstract situation sketched in () is precisely what we explore in the empirical domain of SE constructions. 4 0 See Section for a discussion on Roots in contexts where there is no DP for satisfying their subcategorization features.

240 Mercedes Pujalte and Andrés Saab In the next sections, we first show how our assumptions about feature inheritance derive the basic syntactic properties of SE constructions, and then provide a restrictive way to implement CL-insertion at PF. 4. THE SYNTAX OF SE CONSTRUCTIONS In this section, we present our analysis of the syntactic derivation involved in SE constructions. Our proposal is that anticausatives, passives, and reflexives have the structure in (b), where v lacks j-features and C has (complete or incomplete) j-features; impersonals, in contrast, have the structure in (c), where v is j-complete and C is defective. Let us start with the anticausative alternation, illustrated again in (28): (28) a. La tormenta hundió el barco. the storm sank.sg the ship The storm sank the ship. b. Se hundió el barco. SE sank.sg the ship The ship sank. The crucial property of (28b) is the absence of j-features on v. In this scenario, there is feature inheritance from C to O, creating a C-T-v-O complex. This chain agrees with the IA associated to the Root, valuing its j-features, and assigning a nominative value to the unvalued Case feature of the IA. Crucially, the lack of j-features on v is what blocks the introduction of an EA (see Section ). In (), we represent the relevant part of the derivation, where the bidirectional arrow expresses the Agree relation, and the unidirectional one, feature inheritance: () 4 As for passive SE, we have argued that it essentially involves the same vp structure as that of anticausatives, but the feature encoded on v is not [CAUSE] but [DO] (see Section 2). This means that passive SE can only have an agentive interpretation. Like anticausatives, the IA in sentences like (0) agrees with T

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 2 and receives nominative case. However, passives have a person restriction, as we noted in Section 2. This is shown again in () (cf. () above): (0) El barco se hundió (para evitar más catástrofes). the ship SE sank.sg (for to.prevent further disasters) The ship was sunk (in order to prevent further disasters). () a. Se encontró María. (Ok as reflexive) SE found.sg María Intended meaning: María was found. b. Me encontré yo. (Ok as reflexive) ME found.sg I Intended meaning: I was found. Crucially, the only available reading for the examples in () is the reflexive one. This leads us to the idea that the person constraint in SE constructions is linked to thematic interpretation at LF (see below). The only syntactic difference between passives and reflexives is that the former lack person features. That is, in both cases we have a C-T-v-O complex, but in passives there are no person features on C. Let us see concretely how the derivation proceeds for passive SE. In the first place, v is a fully defective head, which implies that a complex C-T-v-O with unvalued number is formed. As a consequence, the DP agrees in number with T but not in person. (2) 4 The person restriction follows straightforwardly from the derivation in (2), because a DP with person features cannot have its Case valued against defective probes. In this respect, we follow Rodríguez-Mondoñedo (200) with some adaptations who proposes that the complement of existential constructions with haber to have in Spanish is subject to a person restriction due to the fact that v has no person features (in our terms, v is fully defective and inherits a number feature from C-T). He argues that the IA of these existential constructions must lack person features in order to have its Case feature valued. The assumption behind this analysis is that full DPs (but not proper nouns or strong pronouns) can optionally have a person feature as part of their inflectional specification.

242 Mercedes Pujalte and Andrés Saab As for reflexives, they syntactically encode person features. In (4), the derivation for reflexive SE is represented: () Yo me lavo. I ME wash.sg I wash myself. (4) 4 The question is how the IA is also interpreted as the EA and why this is, indeed, the only available reading. We depart from Schäfer (2008) and much of the standard generative tradition in that we assume that no binding configuration is obtained in (4). Instead, we propose that the DP is just interpreted with two thematic roles. The theme/patient role is configurationally assigned and, to some extent, this is also the case for the external theta role. Let us assume then that thematic interpretation takes place in the domain of the vp, as is standard. In a configuration like (4), the IA receives its patient role in the complement position. With respect to the external thematic role, we have an agentive v without a DP in its specifier. We assume now that this thematic role can be discharged to the IA of an active sentence only if v lacks j-features, which is the case with reflexives. Therefore, the IA of a reflexive configuration is accessible for further thematic interpretation. This gives us the following result: () In a vp domain with active-agentive v, the internal DP can be interpreted as the agent of the event only if: (i) v has a non-discharged D feature, (ii) vp is not inherently specified with j-features. It remains to be accounted for why in passive SE constructions the IA cannot be interpreted as the agent of the event. We think that the answer to this Jairo Nunes (p.c.) observes that () seems to be incompatible with the Movement Theory of Control (MTC, Hornstein, and much subsequent work). This is partially true. Indeed, if we are on the right track, movement of the IA to an EA position in the same vp domain should be prevented, otherwise this moved DP could discharge the [D] feature on v and SE should never occur. Nevertheless, movement to thematic positions in control structures is not ruled out by any evident property of our analysis. That is to say, our system is incompatible with a movement theory of reflexivization (see Boeckx, Hornstein & Nunes, 2008; Hornstein, 200), but not with MTC.

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 24 problem has to be linked to the inflectional composition of DPs. Our claim is that to have person features is a necessary condition for being an agent. As we have seen, this is never the case with passive SE constructions. Hence, an arbitrary default interpretation is obtained for the EA at LF. One of the main virtues of the system we are proposing is that it allows us to account for impersonal SE in Spanish with the same mechanism that underlies the other three cases. As mentioned in Section 2, impersonal SE has an agentive meaning and can occur with transitive and unergative verbs (although see Section. for other instances of impersonal SE): (6) a. Se castigó a los culpables (para ganar las elecciones). SE punished.sg to.acc the culprits (for to.win the elections) One punished the culprits (in order to win the elections). b. Se trabaja bien en este lugar. SE works well in this place One works well in this place. We propose that impersonal SE is a case where the C-T complex lacks j-features, but v is j-complete; hence, in (6a) the complex v-o values its j-features against the IA, which, in turn, receives accusative case. Furthermore, the system prevents feature inheritance from v to C (see Section ). Thus, the complex C-T has no features and T gets default rd person singular at PF. In (), we illustrate the syntactic output before Spell-Out. () 4 Just like the other cases discussed in this section, () has a non-discharged D feature associated to v. Finally, with regard to the interpretation of impersonals, given that v is j-complete, the IA receives its patient role and is inaccessible for further thematic interpretation in consonance with (). As in this case v does not have a specifier, LF assigns an arbitrary and generic reading by default in the same way as in passive SE constructions. 2 2 The claim then is that in impersonal SE constructions there is no syntactic representation for the EA (against a long standing assumption; that is, there is no pro (Cinque, 88 and much related work) nor PRO (Mendikoetxea, 2008). See Section. for more discussion.

244 Mercedes Pujalte and Andrés Saab. A PHASE-BASED APPROACH TO CL-INSERTION As we proposed above, the morphology of SE constructions is the overt reflex of the repair strategy that PF performs in order to cancel the D feature on v. In other words, a clitic is inserted post-syntactically whenever a subcategorization feature survives the syntactic derivation. In this section, we propose an explicit mechanism for CL-insertion at PF that determines under which locality conditions this operation can take place. We assume a restrictive theory of the syntax-morphology connection in consonance with DM and, especially, with the framework proposed by Marvin (2002) and Embick (200), according to which morphological operations are sensitive to phases. This approach is directly linked to some recent lines of thinking in the generative enterprise at least since Chomsky (2000, 200). The current debate focuses on what constitutes a syntactic cycle and how much syntactic information is accessible for further syntactic or morphological operations. According to Marantz (200), category-defining heads like n or v are cyclic heads, an idea that Embick formulates in the following way: (8) Category-defining heads n, v, a, and so on, are cyclic heads: such heads define the phases that trigger Spell-Out. (Embick, 200, p. ) 4 Other functional categories like C or D can also be phases, as argued for in several recent approaches to phase theory. In any case, we will focus here only on v, which is also the introducer of EAs, as discussed above. A version of the Phase Impenetrability Condition (PIC) is also adopted in Embick s work. Concretely, the complement domain of a given cyclic head is inaccessible for further PF computation when a higher cyclic head is spelled out. As in other approaches to the PIC, we assume that the head of the phase, its specifier, and adjoined material to the projection of that head are accessible for further computation; that is, edge material including the phase head is active for the higher phase. However, edge material does not include Embick s definition of the PIC is given in (i): (i) Material in the complement of a phase head that has been spelled out is not active in subsequent PF cycles. That is, the complement of a cyclic head x is not present in the PF cycle in which the next higher cyclic y is spelled out. (Embick, 200, p. 4) This definition is connected to Chomsky s (200) weak version of the PIC (see Gallego, 200 for a detailed discussion on weak vs. strong PIC), according to which a given phase head is evaluated only at the next phase level. As noted by Richards (20), the two versions of the PIC are introduced only in relation to the non-phase head T. The point is whether or not T can see inside the vp phase. In Embick s version of the PIC, it is ensured that T can see the O, which seems to be the right empirical result (e.g., tense allomorphy conditioned by arbitrary Roots in English). In the main text, we assume a more basic notion of the PIC (see ) closer to the first version given by Chomsky (2000), where the reference to the next phase head does not seem necessary (see also Chomsky, 2008, p., footnote 24). However, see footnote 8 for more discussion.

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 24 heads incorporated by head movement or by other processes that produce complex heads. In other words, a complex head structure does not modify the cyclic structure of a given phase head. To our knowledge, Marvin (2002) was the first in observing this point. 4 Her formulation of the PIC is given in (): () H and its edge are spelled out at the next (strong) phase. The domain of H is spelled out at the phase of HP. A head h adjoined to H is in the domain of H. (Marvin, 2002, p. 26) The last clause of this definition ensures what we have just observed: head movement does not affect the cyclic structure in which it is involved. Evidently, this is a stipulation that, as far as we can tell, does not follow from any notion of edge, but is forced by empirical facts (cyclic structure within words, incorporation phenomena and head opacity in general; see also Matushansky, 2006). In what follows, we show how this version of phase theory allows us to derive the basic properties of CL-insertion and, at the same time, to make some specific predictions about the timing of CL-insertion... CL-insertion and cyclic heads Recall again the syntactic configuration created by the absence of an EA in Spec,vP (cf. ): (40) After head movement from O to T, 6 a complex head with a surviving D feature on the cyclic head v is formed. Thus, the structure that PF receives has the following form: 4 4 We are grateful to Dave Embick and Pablo Zdrojewski for pointing out Marvin s work to us. What the empirical facts lead us to conclude is that in a complex phase head only the label of this complex head is visible for the computational system. So, for a given phase head Ph, the edge of Ph is Ph 0max (in the sense of Chomsky, ) and non-phase material above Ph 0max. Although this observation does not derive the notion of edge, it seems to have some conceptual motivation. 6 We remain neutral about the syntactic or PF nature of head movement. We are assuming that a dissociated agreement node (see Embick, ) is added at PF as a result of Agree.

246 Mercedes Pujalte and Andrés Saab () As mentioned above, if the D feature on v is not eliminated, the derivation is aborted given (24). In Spanish and other Romance languages, however, a clitic can be inserted in order to satisfy this D feature. What we need now is an explicit mechanism for this process. In other words, we have to give an answer to the question (B) presented in the introduction: (B) How is CL-insertion implemented at PF? We propose that CL-insertion is subject to the same cyclic constraints that apply to other syntactic and morphological processes. In other words, a clitic can be inserted at the top of the complex head T if and only if the surviving subcategorization feature is visible in the cycle in which the clitic can be inserted. As shown in (), this is exactly the case. In effect, by the PIC in (), v, as the head of the phase, is accessible for further computation, although material in its complement is not: (42) A very important prediction of this system is that unsatisfied categorial features on Os are not visible for CL-insertion because they are located in the complement of the phase head. Thus, our system captures in a principled way the general observation that expletive-like elements are not attested for IAs. Only categorial features on T or v are subject to PF-repair in the sense we are proposing. As an illustration, consider a situation in which a O with a D feature is not merged with a DP in the syntax. Under this scenario, O [D] is sent to PF with its D feature not discharged. Once in PF, CL-insertion is impossible because the D feature is in the complement domain of v and, consequently, invisible for further computation. 8 4 8 Jairo Nunes (p.c.) and an anonymous reviewer wonder whether the same facts can be captured under a less restrictive version of the PIC (say, Embick s version in footnote ). We think that the answer to this question depends on several assumptions related to the position of

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 24 (4) 4 This approach explains why null objects with arbitrary reference in Spanish and other Romance languages (see Rizzi, 86 and much subsequent work) never occur with overt clitics, but implicit EAs like the ones we are addressing here do. Notice, first, that null objects are not allowed under normal circumstances: (44) a. Juan compró. Juan bought.sg b. Juan hace. Juan makes morphological clitics and to their phase nature. If clitics are inserted in the C-domain, then the clitic cannot see the Root under that version of the PIC (and, of course, under any current version of the PIC). But assume that clitics are indeed inserted on T. Then, it seems that Root properties are visible in the T domain, unless clitics are cyclic nodes (i.e., D nodes). If this is the case, a clitic can never see the Root regardless of whether or not it is adjoined to T. Indeed, the prediction is more general. Consider the case of impersonal rd person plural in Romance. As noted by Cinque (88, p. 4), arbitrary readings with null rd person plurals are only available with transitive or unergative verbs, but never with unaccusatives: (i) Lo encontraron muerto. him found.pl dead They found him dead. (ii) Telefonean. telephoned.pl They telephoned. (iii) Llegaron temprano. (OK under the referential reading) arrived.pl early They arrived early. Under our analysis, the ungrammaticality in (iii) follows straightforwardly because of the PIC. This means that plural marking in these contexts is also triggered by an unsatisfied categorial feature related to the EA. Notice however that the arbitrary SE/PLURAL alternation is not semantically vacuous: the sentence in (ii) for instance cannot be uttered in the same context as an impersonal SE sentence. We conjecture that categorial features have to be decomposed in different subtypes with different semantic import. We cannot offer here a semantic analysis of arbitrary readings, but we would like to stress that the contrast in (i)/(ii) and (iii) which is not easily derived under current assumptions gives additional support to a phase approach as the one we are pursuing here. We are grateful to Jairo Nunes (p.c.) for drawing our attention to this point.

248 Mercedes Pujalte and Andrés Saab 4 c. Juan corta. Juan cuts However, null objects are attested in contexts where a human and generic reading is available, as noted by Rizzi (86). Some Spanish examples are given in (4): (4) a. Los fantasmas asustan. the ghosts frighten.pl b. La lectura ayuda. the reading helps c. El psicoanálisis cura. the psychoanalysis cures These examples seem to form a minimal pair with impersonal SE constructions, where a similar generic and human reading is obtained but where, at the same time, CL-insertion is mandatory (cf. 6b): (46) Se trabaja bien en este lugar. SE works well in this place One works well in this place. This absence versus presence of CL-insertion in IA and EA positions is a specific prediction of our system, as we have already noted. Nevertheless, it remains to be answered how the examples in (4) are derived. One possible analysis is suggested by Rizzi (86): null arbitrary objects in cases like (4) are pro licensed and identified by V. If this is the case, pro just merges with O and discharges the D feature on O. Another possible alternative could be that O has no D feature on it from the starting point of the derivation and, as a consequence, no DP merges with it. Under this scenario, no subcategorization conflict arises and the Encyclopedia is in charge of the interpretation of the O domain (see Embick & Marantz, 2008 and the references therein). In other words, the sentences in (4) and (46) are interpreted via different underlying mechanism. 20 20 That the interpretative mechanism for this type of implicit objects is different from the interpretation of arbitrary SE seems to be empirically correct. In a sentence like (i), the implicit object can be interpreted either as animate or inanimate. What determines this type of ambiguity is the particular semantics of the Root. If the interpretation of the implicit object depended on a default mechanism at LF like in impersonal SE contexts (see Section 4), then this ambiguity would not be derived. (i) Ese cuchillo corta. that knife cuts Other particular interpretations of implicit objects can also be derived from the Encyclopedia (e.g., Juan toma, John drinks meaning that Juan drinks alcohol).

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish 24 4 It should be possible to distinguish these alternatives on empirical grounds. In principle, both raise the question of why arbitrary pro insertion or the lack of a D feature on v is not possible in EA position. If any of these options were possible, we would expect impersonal SE without SE in finite clauses in Spanish, contrary to fact: (4) a. Aquí (se) puede fumar. here (SE) can.sg smoke b. (Se) castiga a los culpables. (SE) punishes to.acc the culprits However, this could be a hasty conclusion both for empirical and conceptual reasons. Consider the nature of real passive sentences, which, as is well known, have an implicit agentive meaning, as the presence of by-phrases shows: (48) Juan fue reprimido por la policía. Juan was repressed by the police Let us then assume that the structure of an analytic passive is indeed a case where vp does not have a D feature on v, an option that, by hypothesis, is linked to the absence of inflectional features on v. That is to say, fully defective v without a D feature is an analytic passive. The licensing of by-phrases in passives can be related to this particular configuration straightforwardly if it is the case that the presence of a D feature on v is incompatible with an agentive by-phrase. Absence of categorial specification on v, on the other hand, leaves room for an agent to be realized as a PP. If this is correct, we can also explain why by-phrases are not licensed in contexts of passive/impersonal SE sentences: (4) a. Se reprimió a los maestros por la policía. SE repressed.sg to.acc the teachers by the police b. Se destruyeron los puentes por el enemigo. SE destroyed.pl the bridges by the enemy This tentative conjecture about analytic passives is exactly what is expected on our assumptions about inflectional and subcategorization features. So far, we have assumed that C or v can be selected from the numeration without j-features. There is no reason, we think, why this cannot also be the case with subcategorization features. At the same time, this idea leads us to conclude that argument structure alternations and, specifically, the passive alternation in all its variants are just an epiphenomenon that results from the possible combinations of inflectional and subcategorization features. Thus, there is not a special voice feature that gives us the different possible semantic and syntactic expressions related to voice; instead, what we have is different degrees of formal

0 Mercedes Pujalte and Andrés Saab feature realization on v. As for v [DO], we have the following combination in the vp domain: Transitive: Passive/reflexive SE; Analytic passive: (0) a. b. c. 4 Whenever C-T is j-complete, (0a) is a typical transitive structure. (0b) is the passive SE structure, if C-T is person-defective, or the reflexive structure, if C-T is j-complete. (0b) can also be an instance of the anticausative SE when v is [CAUSE]. Finally, (0c) is the analytic passive, the most underspecified version of v. This approach to the extent that it captures broad generalizations on argument alternation favors an analysis of the null generic objects in sentences like (4) in terms of absence of subcategorization features on O, instead of an arbitrary pro in object position. Before concluding this section, we should complete the picture by mentioning another specific prediction of our system: analytic passives cannot occur with arbitrary SE. There are two reasons for this. First, a D feature cannot occur on O for the locality reasons previously adduced. Second, if the D feature is inserted on v, then we obtain the passive or impersonal SE, not the analytic passive. Nevertheless, this prediction seems to be falsified by empirical data: analytic passives do occur in impersonal SE contexts: () a. Cuando se es condenado por los jueces,... when SE is sentenced by the judges When one is sentenced by the judges... b. Aquí, se es reprimido por la policía. here SE is repressed by the police Here, one is repressed by the police... As the compatibility with by-phrases shows, these sentences are real cases of analytic passives. However, we believe that examples like () do not constitute counterevidence to our system; rather, they are exactly what we predict once they are explored more carefully. What should be kept in mind is that our system only prevents D from being associated with the Root whenever there is no DP merging with it in the syntax. Moreover, we predict that an analytic passive cannot have an underlying v [D]. Yet, there is another alternative namely, it could be the case that T is associated with a D feature. Plainly, this is just the so-called EPP-property of T (see, for instance, Chomsky, ). Therefore, the underlying representation for analytic passives with SE should

Syncretism as PF-Repair: The Case of SE-Insertion in Spanish be like in (2), where T has a D feature that is satisfied by merging the clitic at PF in the way previously explained: (2) 4 We believe that this is a good result both on conceptual and empirical grounds. Conceptually, our system leaves room for a more general theory of expletive insertion that can be extended to well-known cases of expletive insertion in languages like English. On the empirical side, we expect impersonal SE constructions to behave differently intra- and cross-linguistically. Regarding Romance languages, it is well known that not every impersonal SE behaves in the same way (see Cinque, 88; Dobrovie-Sorin, 8; D Alessandro, 200, among many others). This is also the case with impersonal SE in the Spanish grammar. Crucially, sentences like () do not pattern exactly like impersonal SE in transitive contexts. A clear piece of evidence is that they have a different distribution in non-finite clauses, as first noted by Cinque (88) for Italian (see also Dobrovie-Sorin, 8): whereas impersonal SE in transitives can occur in absolute clauses, analytic passives with SE cannot. In this respect, compare the following minimal pair: () a. Al castigarse a los culpables... to.the to.punish.se to.acc the culprits Punishing the culprits... b. Al serse castigado... to.the to.be.se punished One being punished... Cinque stipulates that there are two kinds of impersonal and nominative SE, a [þarg] SE and a [-arg] one. Argumental SE underlies transitive and unergative SE constructions, whereas non-argumental SE is instantiated in passive and unaccusative verbs. According to Cinque, only argumental SE can occur in non-finite clauses (see Dobrovie-Sorin, 8 for criticism and an alternative analysis). In our analysis, this stipulation is no longer needed. Let us see why. Pérez Jiménez (2008) has presented compelling evidence in favor of the idea that absolute clauses have to be analyzed as a case of predicate fronting to